20VC: Foundation Models: Who Wins & Who Loses | How Economies and Labour Markets Need to Change in a World of AI | China vs the US in an AI Race: What You Need to Know | Rich Socher, Founder @ You.com

Rich Socher is the Founder and CEO of You.com. Richard previously served as the Chief Scientist and EVP at Salesforce. Before that, Richard was the CEO/CTO of the AI startup MetaMind, which Salesforce acquired in 2016. He is widely recognised as having brought neural networks into the field of natural language processing, inventing the most widely used word vectors, contextual vectors and prompt engineering. He has over 150,000 citations and served as an adjunct professor in the computer science department at Stanford. In Today’s Episode We Discuss: 04:10 Winners & Losers: OpenAI, Gemini, Claude 08:59 How Partnerships Could Decide the Winners in AI 12:42 China vs US: Who Wins the War for AI 25:50 How Society and Economics Needs to Change in a World of AI 34:04 What Jobs Will Be Replaced, What Will Not 36:04 How Europe Needs to Change It’s Approach to AI 41:06 How AI Will Change Health and Longevity 43:10 AI in Consumer and Enterprise Markets 49:30 Quantum Computing and AI Misconceptions 56:57 Longevity, Personal Reflections, and Future Outlook

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Harry Stebbings?

  You are listening to 20 VC with me, Harry Stebbings. Now, what on earth is going on in the foundational model layer? It seems like there are new releases, new updates, new winners and losers every week. I wanted to sit down with one of the best in the space to understand what really is going on and how we should think about it. So joining me in the hot seat today, we have Rich Socher, founder and CEO of U.com. Before founding U, Rich served as Chief Scientist and EVP at Salesforce. Before that, he was the CEO/CTO of the AI startup MetaMind, which Salesforce acquired. He's also widely recognised as having brought neural networks into the field of natural language processing, inventing the most widely used word vectors, contextual vectors and prompt engineering.

• 0:42Harry Stebbings?

  No one better for this topic today, and it was a fantastic discussion to analyse where we are with foundational models. But before we dive into the show today, SecureFrame empowers businesses to build trust with customers by simplifying information security and compliance through AI and automation. Thousands of fast-growing businesses, including NASDAQ, Angel List, Doodle and Coda, trust SecureFrame to expedite their compliance journey for global security and privacy standards, such as SOC 2 and ISO 27001, CMMC, MIST standards and more. Backed by top-tier investors and corporations like Google and Kleiner Perkins, the company is among Forbes' list of the top 100 startup employers for 2024. G2's Best Software Awards for highest satisfaction products

• 1:27Harry Stebbings?

  and a recipient of the 2024 Cyber Security Excellence Awards, something I definitely never got in school myself. Learn more today at secureframe.com. And speaking of trust and intelligence, let me tell you about Harmonic. Did you know that half of the 27 companies started last year by OpenAI alumni are still in stealth? I discovered this on Harmonic, the complete startup database used by Excel, Insight, Menlo and hundreds of other leading VCs, as well as go-to-market teams from the likes of Notion, Brex and Google, to find the best startups and founders, even in stealth. Few things annoy me more than missing a round where we know the founder and should have led the round. But now, this is a problem of the past. Harmonic maps everyone your team

• 2:10Harry Stebbings?

  has ever met, emailed or connected with to the source of truth for startups. So when that company that was just a little too early suddenly gains traction, you won't be too late. Or when you find the perfect company for your thesis or product, you'll know that Sarah just happens to know their CTO. How does Sarah know everyone? Who knows? She's a genius. Learn how VCs and GTM teams find the best startups six months ahead of the competition on harmonic.ai. And while Harmonic helps you stay ahead of the curve in startup investing, what if your phone could help you do the same with your wallet? We spend nearly half of our waking lives glued to our phones, upwards of 50 hours every week. Recently, one company transforming this reality stood out so much I personally became a shareholder.

• 2:53Harry Stebbings?

  Mode Mobile. Mode Mobile created the Earn Phone, a smartphone that pays you for daily activities. Instead of big tech profiting billions from our attention, Mode returns over 325 million directly to users through earnings and savings. Mode's revenue surged an incredible 32,000% in three years. Recognized by Deloitte as 2023's fastest growing software company in North America. And here's why I'm excited. Mode's equity offerings have raised over 30 million from 20,000-plus retail investors, one of 2025's standout public raises. You can now join me as a shareholder with as little as $1,000 at invest-modemobile.com/20vc for a limited time. Unlock up to 100% bonus shares and a free Earn Phone.

• 3:38Harry Stebbings?

  Email us for investor brief at [email protected] or check out invest-modemobile.com/20vc. You have now arrived at your destination. Richard, dude, I am so excited for this. We did one before, remote. It is so much better in person. And so when you said you're in London, I was like, we have to make this happen. So thank you for joining me.

• 3:59Harry Stebbings?

  Thanks for having me. It's a beautiful day.

• 4:00Harry Stebbings?

  It is a beautiful day. It's never like this. I want to start with a little bit on you for those that are listening for the first time. High level on why you're a rock star and what you is.

• 4:09Speaker 02

  High level, did my PhD at Stanford and credited for having brought neural networks into the field of natural language processing. It was a very controversial idea at the time. It's very obvious in retrospect. It's just sort of a story of my life. So, you know, brought like word vectors, like improved those massively and sort of built one of the two most popular word vectors. Then we pushed contextual vectors, so you can pre-train not just a single word vector, but a whole sentence embedding. And that then became LMO, which became BERT, which is one of the most cited papers still in the world. And then we invented prompt engineering, which was majorly rejected publicly on Open Review. An idea that made no sense to the reviewers. And now in retrospect, it's so obvious, like no one could have even invented it.

• 4:53Speaker 02

  It's just like, of course you can ask questions to one model and no matter the question, you'll get an answer. So done a lot of research after PhD, did a startup called MetaMind, was acquired by Salesforce, where I became chief scientist. After four and a half happy years at Salesforce, I started U.com. U.com basically came from this idea that we have a single model now, a single neural net that can answer all the different kinds of questions. So clearly, people on the internet should get better answers than lists of blue links that we get from Google. And so we started with that premise, but eventually realized a lot of people ask fairly simple questions to Google. Like what's the weather tomorrow? What's the score of the soccer game? What's the, who's the president of the US? What's the price of the stock?

• 5:35Speaker 02

  And like, on a lot of those questions, you don't really have the opportunity to be 10x better than a Google. You get that answer within one second and that's it. And so we realized eventually the killer app for large language models and complex answers is an enterprise. And so we're now helping companies with answers, agents and a path towards AGI at U.com.

• 5:57Harry Stebbings?

  I would love to start with top level because it is very noisy and it is very confusing to understand what's going on. When you evaluate where we are in the LLM landscape today, how do you evaluate the current status of where we are?

• 6:11Speaker 02

  Boy, I think AI is kind of this tide that's rising. But on top of that tide, you have a lot of little hype bubbles that come up and down. And you know, in some ways, you can think about when Sam, for instance, says the next generation of models will be as good as a PhD student. The corollary there is that most jobs don't even require a PhD. If you do service for DoorDash or something, like you don't need a PhD to answer service questions. And so LMs are already good enough. They just need to be brought into companies to be actually made useful. And so I think that's sort of one state. And then the future state is, of course, we will get even better at reasoning. And at some point, there are very, very narrow niches where these models can be as

• 6:56Speaker 02

  good or better than an expert human. So there's still a lot of room to grow. And so in fact, it's such a confusing state that part of this book I'm writing includes a chapter on what I call the upper bounds of intelligence, where I essentially group intelligence into 10 different dimensions or groups of dimensions. And then we can kind of say, in this dimension, there is like a fairly low upper bound and we're fairly close to it. For example, object detection and computer vision. It's actually kind of solved. We can classify most objects on the planet now. And the upper bound that that type of intelligence can ever get to is all objects on the planet. And so I think we're already 80, 90% there.

• 7:38Speaker 02

  But in the other bounds, like knowledge, well, if you include the molecular composition of every planet in the universe as part of knowledge, we are astronomically, quite literally and figuratively speaking, away from ever having AI reach that upper bound that is basically in the world of physics. And so the state of LMs is very perplexing right now.

• 7:59Harry Stebbings?

  We were talking about the speed with which the landscape moves. We are seeing this seemingly intense commoditization of LLMs. Do you agree with the commoditization of LLMs?

• 8:08Speaker 02

  Hundred percent, yeah.

• 8:10Harry Stebbings?

  If they are being commoditized, is there value in them?

• 8:13Speaker 02

  It's very interesting, but on value, you have to also differentiate between value creation and value capture. I think LLM Companies, especially just the pure, thin infrastructure layer of L L Ms are going to look, I think, more and more like telcos in the sense that it's high capex, huge expenditure to build it, especially if you want to build it from scratch. It creates a lot of value in the world. You can't build an Uber app if people don't have internet everywhere, but you don't necessarily capture that value, just like Vodafone and T Mobile and whoever don't get a cut of Uber working now, right? And so I think that's kind of the mental model I built for L L Ms.

• 8:59Harry Stebbings?

  Can I interject there and say the core of a telco business model is actually sustainability and retention? It's why they often tie you into very long term contracts. If you look at the L T Vs of a Vodafone or a T Mobile customer, these are incredibly long, ten year plus.

• 9:13Speaker 02

  You're right, that's where it breaks. It's even worse because the one player it breaks is software. So you don't have, you have even less of a moat and with open source, it's even more.

• 9:21Harry Stebbings?

  Now, I said this to Kevin Scott at Microsoft and he said, what's the moat with search? Go from Google to Bing, just the same, but you don't, you stay with Google. And he didn't obviously say that because he's obviously with Bing. But my question to you then is like, how do you see the distribution of value across the L L M space with the recognition of that?

• 9:40Speaker 02

  That's why I said, if you're in that thin infrastructure layer, and that was an important qualification, because OpenAI is a consumer app company. They make their revenue, the vast majority of their revenue from a consumer app called Chat G P T. If you're now just in that API infrastructure layer, it's very different. You have a lot more pressure. Anthropic has a lot more pressure to keep building the best models because Claude is so much smaller in terms of market share for the consumer app. And so that's why that analogy doesn't work. And you're right, like consumers, once you're really famous and you cross that threshold of just like being well known, being the default for a lot of people, all the other L M apps companies are almost rounding errors to Chat G P T.

• 10:21Harry Stebbings?

  Why do Anthropic still keep the consumer products of Claude? They are clearly very, very good at engineering. They are clearly used by the best engineering. Companies in the infrastructure layer, why bother keeping the consumer facing products?

• 10:35Speaker 02

  I think in AI, if you say I have the best model, but people can play around with it very quickly, people call BS on you.

• 10:41Harry Stebbings?

  But there's not a data network effect, there's not a data acquisition play really there on user inputs that make models better.

• 10:48Speaker 02

  It could be, it could be that they use random conversations and maybe train and guide it.

• 10:52Harry Stebbings?

  Do you think we're seeing the specialization of these different providers? Like you said there, you have Anthropic obviously very much focusing on engineering, so to speak. The consumer product of ChatGPT, you said about you really focusing on enterprise and enterprise workflows. Are we seeing this realization of shit? There's no inherent value in horizontal products, we need to be specialized.

• 11:10Speaker 02

  Maybe, just because ChatGPT owns the consumer market that hard, you have to find a different part of the niche. And we're saying, we're seeing that play out in a variety of different ways. Like, we're also focusing more on enterprise. My hunch is other people will follow us into that, because that is just like normal consumers again. Either they have a majority very simple questions, or they don't want to pay. Ads and chat are really hard, we actually evaluated that. They work about ten to hundred X worse than search ads, and you have twice the cost about.

• 11:39Harry Stebbings?

  Talk to me about that, sorry. So ads in LLMs do not work?

• 11:43Speaker 02

  They work ten to hundred X worse than chat search ads. So this is in search, right? You like, Google at some point found that no matter how bad they make the product, people don't know how, where else to Google. Right? So they, they default. Also, this scary and sad statistic is like eighty percent of all iPhone users never change a single setting of any kind. Whatever is the default gets used, and that's why like Google pays Apple twenty billion dollars a year to be that default, right? And so it's very hard to move away from that kind of powerful of a lock-in. And then if you realize that, you can say, well, if we need to make more revenue this quarter compared to last quarter, let's just do six ads instead of five ads. And when the organic link results get worse and worse, because they're SEO, and then everyone knows they're kind of getting terrible.

• 12:28Speaker 02

  Turns out you make even more money, because the product gets worse, but the ads become more and more relevant. And so after doing that for ten years as an untouchable monopoly, the experience suffered, right? And so there is like some potential here, but the default is still very, very strong.

• 12:42Harry Stebbings?

  What happens from this point on? We had like Guillermo Rausch from Vessel tweet the other day that actually they've seen conversions to their site transfer 1.6% to now 4.5%. It used to be that that's like direct from ChatGPT. My question to you is like, do we just see this slow migration away from Google? Do we see Gemini integrated into Google search? Well, how do we see that conversion happen?

• 13:07Speaker 02

  There's sort of big waves in consumer applications of bundling and unbundling. For a while, I was hoping we would be in a bundling kind of wave still, but I think it's fairly clear that we're still in a very large wave of unbundling. Consumers are okay going to Yelp for a restaurant review and then going to if they really care about the weather because they are into flying sports or something, going to a specific weather app like Wendy. Then they go to a specific app like Uber Eats to like get food delivery or DoorDash or whatever to get their food. So if they want to make a very small purchase, that's like 20 bucks and they don't care about it, they search directly on Amazon.

• 13:46Speaker 02

  Young people now look directly on TikTok because they want the food to look good in their Instagram or TikTok or whatever videos. And so I think there's a huge unbundling wave. And so LM's as part of that unbundling wave of Google, LM's will capture whenever you have a more complex question. And then you just look at like how many people have complex questions in their lives and the more you are a knowledge worker, the more complex questions you have in your life. And they're often you have them at work and that's where like efficiency matters too. And those are some of the many reasons why you dot com moved into enterprise.

• 14:21Harry Stebbings?

  Does one need to be the best and innovative today when you can just distill very effectively?

• 14:27Speaker 02

  I think that answer depends on the dimension that you want to be best at. I do think it's good to, it's obviously always helpful to be the best. We pride ourselves to be the most accurate and that is a never ending game. So whenever you say you're the best, you're the best in that moment and you have to keep working on it. And it does help, does help with marketing and branding and sales for the most part. I feel like we have still not maxed out our abilities on the marketing side and branding side of things. But at least sales works well enough now that we're really increasing revenue and that's ultimately what matters.

• 14:57Harry Stebbings?

  Do you think we overestimate adoption in the short term for large enterprises and underestimate it in the long term?

• 15:03Speaker 02

  We have seen, actually, some of our largest enterprise customers now had failed OpenAI projects or had failed.

• 15:13Harry Stebbings?

  Why didn't they fail OpenAI projects?

• 15:15Speaker 02

  I'll tell you that, or built their own with some APIs. Those are some of the largest customers we've had, and they fail usually for two reasons. One is adoption. They had to pay a thousand seat licenses for OpenAI, and then six months later, they realize only 6% are actually using them every week. So they've been sitting and paying for 90% licenses.

• 15:36Harry Stebbings?

  Why is that?

• 15:37Speaker 02

  With AI, every person will become a manager. But most people are not used to managing other people or processes. They're used to being individual contributors, doing a specific type of work and doing it well. Now, when you become a manager, you have to learn to distill all your knowledge in a very succinct and unambiguous way to another entity, in this case, an agent, right? And so we're helping people essentially train up their own agents. So whatever process they have in their company takes them enough hours, they have, they repeat it every couple of weeks or every couple of days. Like, we teach them on how to actually tell that to an AI, and then they just have their own agent, and it'll just automate that for them.

• 16:16Harry Stebbings?

  Do you think we will operate in a world of horizontal agents that do many different tasks, and it's kind of like your personal assistant? Or do you think we operate in a world of many verticalized agents, which specialize for very specific tasks?

• 16:28Speaker 02

  Yeah, I think a lot of companies right now would love to immediately own the end user and then do all of that below. But my hunch is, we've gone through this at U.com too, when we're still doing more consumer, we're looking into like, basically, we wanted to make it easier for users to get things done, right? This is the same idea that now we see with agents. And when you look into it, like, I remember this, this demo where, where a startup founder was like, had his little device, and he's like, I want to have a trip to London with my four kids and, you know, on this weekend, and then one, two, three, and it's done. And I was like, that was definitely BS. Like, there's no way that was true, because when you've, if you've ever booked a trip, right, there's so many little nuances. And you realize, like, as much as I love natural language, natural language is not the single.

• 17:13Speaker 02

  Best interface for a lot of different types of answers. Sometimes you want to see a map, like sometimes you want to see a table. Sometimes you want to see a map with a bunch of specific overlays. And sometimes it's okay to just do it over voice. But even flight booking becomes complicated and changes over time. Let's take something as simple as booking a flight when you're a student. And then six months later, after you've had a job. Well, once you have a job and you have more money but less time, you'd rather pay extra for a direct flight versus a one stop flight. The AI needs to know all these subtle details about you to get really good. And we're sort of in this valley of disillusionment on a lot of these, what I call action agents that go on the web and actually do something for you and take actions that you can't undo. And say you buy a ticket that's not refundable or something.

• 17:59Speaker 02

  There's this valley of disillusionment that we're in right now because the agents just don't know enough yet about the user.

• 18:06Harry Stebbings?

  Do we have the data to allow these agents to operate effectively? A lot of internal processes and actions within companies aren't actually codified in data. Do we actually have that?

• 18:17Speaker 02

  In some consumer use cases, you would assume if Google really wanted to have it, Apple really wanted to have it, in theory, they could.

• 18:25Harry Stebbings?

  You've been quite vocal about Google before. Are you impressed by their latest enhancements with Gemini? And where they've really positioned themselves?

• 18:33Speaker 02

  I think it's never been a question of technical strength for Google. It's just a question of classic innovator's dilemma. You make money by showing ads and lists of blue links. So it's hard to give people just a straightforward, useful answer. And they have to play around with that now because there's too much pressure. But in a perfect world for them, not perfect for the end user, they wouldn't want to change that that experience. It just prints $500 million a day. Google made this so much money, they didn't know what to do with it. They don't want to pay dividends because then you kind of admit defeat that you can't grow anymore and you don't know, you ran out of ideas. So you have to keep doing something. And they built internet balloons and self-driving cars and like, you know, just.

• 19:16Speaker 02

  Infrastructure, fiber infrastructure, they have so much money, you don't know what

• 19:19Harry Stebbings?

  It's also very difficult to acquire because you have the regulatory provisions, which mean it's super, super hard to get anything through that's non-core, which is why you get something like Wiz, I think, which is like enterprise is not core enough for it to be blockable by any regulators. So I totally agree with you there. But I do think Google are actually best positioned. When you think about one of the most important things being touchpoints to end consumers, I don't think anyone other than them and Microsoft are actually better positioned.

• 19:43Speaker 02

  I mean, in theory, Apple would be so well positioned, but in practice, they've been not doing much.

• 19:48Harry Stebbings?

  What is the thing that's holding back the progression today of L L Ms and specifically how we use them?

• 19:55Speaker 02

  I think one is that personalization, and that's why it's so easy for people to switch around L L Ms too, you know, like DeepSeek overtook almost every other thing other than Chat G P T within the weeks, like with almost no proper like sort of marketing, like other companies in our space spend millions of dollars every week on marketing, right? And like DeepSeek comes in and

• 20:15Harry Stebbings?

  What do you think they did to do that?

• 20:17Speaker 02

  I mean, obviously, one, the fact that it was the first open source model that wasn't just almost up to par, but like in some cases was actually above the closed source models was just a shock, right? You could almost hear billions of dollars of V C investment evaporate sort of into the ether when that happened. Everyone said that should have been impossible. And there's a narrative of you've got to have billions and billions of dollars to be able to even compete. So don't even try. That was the big shock. The timing was great, too. Like, it was actually for several weeks, it was the absolute best model, which is harder and harder, you know, like this weekend, we have a new Llama four model. Why was it launched on a weekend? Maybe they know something we don't. Maybe in like a week or two, there's an even better model that's going to come

• 21:01Speaker 02

  out, right? There's some luck involved, too. And then just like that the fact that it came out of China gave it even more press, controversial that that could have happened in China. And of course, like, why are you surprised? China has been incredibly good at taking technology and making it scalable and cheaper.

• 21:18Harry Stebbings?

  One of the big elements you mentioned there was the cost effectiveness with which they said they trained the models. Do you buy how efficient they were in when they stated how much money it cost?

• 21:26Speaker 02

  Or not, like I think they had, and that was part of their marketing narrative and part of the other marketing narrative for a closed source company because it's super expensive and they all have sort of underlying reasons of why they pushed the number to be super high or why they pushed the number to be super low. It's also clear that it probably cost them a hundred, two hundred million dollars, but it's still incredibly cheaper than billions of dollars that we're told it would take to to train these kinds of models. So the fact that, you know, I think they floated like five, six million dollars, that was maybe the very last training run at best. If you just count electricity costs or something, but you know, electricity costs can be higher and they they fluctuate like buying the GPUs is not included in that.

• 22:06Speaker 02

  Generally in model development, you train one finally like really good, final good model. On the path to that, you had to run many what we call ablations or hyper parameter runs where you tune a little bit like, should this joint be, you know, moving this much or that much, right? Like these large models have thousands of these hyper parameters and you need to tune them. And so on the path towards the best model, you usually train hundreds, if not thousands of smaller models with increasing sizes. So all of these will have cost like several millions of dollars too, or maybe hundreds of thousands and they're smaller. And so, long story short, you put all that together, it was probably closer to a hundred, two hundred.

• 22:42Harry Stebbings?

  Is distillation wrong?

• 22:44Speaker 02

  No, it's a useful thing to make models smaller and yeah.

• 22:48Harry Stebbings?

  Do you think models are venture investments? When you look at the dilution that happens to venture investors over the successive rounds, it costs so much money to do. And then you also have stock-based comp, which makes it just even more dilutive. I don't think any venture investor is going to make money from L L Ms.

• 23:04Speaker 02

  Yeah, again, if you're just in that narrow niche of infrastructure, rather than you own the end user in some capacity, you own some like vertical. I have personally stayed away from investing in any pure L M infrastructure companies, yeah.

• 23:19Harry Stebbings?

  If you're an investor today full time, where would you be spending the most time?

• 23:22Speaker 02

  Early stage, strong technical teams that also have some market insights into a specific vertical, into a specific app. And one of the big verticals that I love so much and think from first principles is the right time to buy right now is in biology too. And so biotech is essentially a perfect storm. Right now, the markets are really down, public bio. Valuations are much lower for early-stage startups, even if they already make good revenue. And the technology is just the perfect tool to really push biology to the next level, so.

• 23:53Harry Stebbings?

  You said there about kind of reasonable pricing with bio. Do you think we're in a bubble in terms of AI early stage?

• 24:00Speaker 02

  Again, this is sort of where I think the analogy is, the tide is rising, but it's also hot. So from time to time, there are bubbles on top of it. I don't think we're in an AI bubble, period. I think intelligence, the fact that the marginal cost of intelligence goes down is the same as like the marginal cost of electricity or coal or something going down, but we will just use it more and more. And like a couple of years ago, I tweeted and read about this thing called Jevons paradox. A couple of like weeks or months ago, some other people have found it also and talk about it a lot. But it is, for those who haven't yet seen it, it's a very useful analogy here. So in the first industrial revolution, like eighteen sixties or so, Jevons was an economist and he looked into the price of coal.

• 24:43Speaker 02

  And a lot of the smartest engineers and minds at the time made more and more efficient coal, like steam engines. And so he eventually thought, and a lot of people thought, well, if steam engines get more and more efficient, then we'll need less and less coal. So the price of coal will go down. But what actually happened is you just use steam engines in more and more places. Eventually, you know, they could create electricity, they can move steamboats and so on, and you just used it more and more. And so the price of coal actually went up instead. And so I think the analogy here is that, yes, AI, like the cost, the marginal cost of intelligence keeps going down. But that just means we're going to use it more and more places.

• 25:19Harry Stebbings?

  Where do you think we don't use it today, or don't even think to use it today, where we will be using it?

• 25:24Speaker 02

  Two areas where I think almost every person also agrees on the planet that it's not about. The jobs, but about the outcomes and that's research and medicine. Most people don't say, I want more PhD students like in the world. They just want the cool things that PhD students develop. And most people don't say, I want more jobs in healthcare. They just want cheaper, faster, better healthcare. And so I think those are two beautiful areas that still were. They haven't had their chat GPT moment and I'll maybe mention one more, which is AI and economics. Writing this book right now, we've had this paper in twenty eighteen, I think. Called the AI Economist and the field of economics hasn't had their chat GPT moment yet. Cause it's such an old, slow moving field.

• 26:06Speaker 02

  They don't have archive papers, they don't have conferences where you publish every couple of months. And so we actually built this two level reinforcement learning model. Where you had an AI economist and you had a bunch of intelligent agents that were just maximizing their own utilities. And they're adapting to different taxation schemes and tariffs and subsidies and so on. And you basically could now ask this model, the system. What's the most efficient taxation and subsidization scheme? For you know, maximizing this objective that I'm giving you. And you know, people can disagree on what the objectives are. One reasonable one was productivity multiplied with equality. And then that system could simulate billions and billions of years of taxation and subsidization schemes.

• 26:49Speaker 02

  Until you basically would find the optimal brackets and whatnot. And just imagine how many millions of lives were like lost to humanity, trying to figure those things out. And instead we could ask an AI to give us some advice on like, what would be the actual best setup? If you have a certain set of goals. You wouldn't have to like try out like a massive tariff change in the world. You could just like ask an AI model first to simulate it for you. That is an underrated area of AI.

• 27:18Harry Stebbings?

  That's fascinating. I've never heard about that before.

• 27:21Speaker 02

  That paper, we did that work at Salesforce and the mark like the science marketing. You realize like even science, you think, oh, it's just like some of one has a eureka moment. And then because of that, they like become super famous and everyone loves that idea. Science is also a human system and you have to do brand and marketing. Like something that DeepMind in London here does incredibly well. Their their science marketing is is probably the best in the world. And so that paper never quite had its moment in the sun yet.

• 27:48Harry Stebbings?

  One thing I worry about is like the excitement around robotics. I find that robotics have not had their chat GPT moment on the foundation model side. How do you think about robotics as having their chat GPT moment or lack of yet? And maybe excitement or lack of excitement that you have towards it moving forward?

• 28:04Speaker 02

  That is a great question. I think the tricky bit in robotics is that a part of why chat GPT had this amazing moment is that it's so general, right? You can just ask it anything. So the equivalent to the generality of a chat GPT is a humanoid, but the humanoids don't work really well yet. The problem is when robotics, like the cost, right? You only wanna build specific types of robots when there's a highly scalable process. And now, once there's a highly scalable process, the humanoid form factor is not the most optimal factor, right? Like if you have like a highly scalable process is in the fields, in agriculture. You don't want a bunch of humanoid robots hunching over and like weeding. No, you just get a frigging like

• 28:45Speaker 02

  massive tractor with a bunch of lasers and like thousands of little arms and spray like cannons. And then you just like either zap the weeds away. You don't have them hand plucked by like a humanoid robot, right? And so for almost every process that is highly repeatable, there's a better, more quickly evolved new hardware form than five fingers on two arms. And so.

• 29:09Harry Stebbings?

  And so humanoid only works in cases of ambiguity?

• 29:12Speaker 02

  In ambiguity and where you have a massive scale of many different tasks. In environments that were custom built for people, where maybe the speed and efficiency doesn't matter as much and so on. Like just take like even for consumer, right? You have a dishwasher. You don't like, imagine you could have a humanoid doing the dishes, but it would be so much less efficient than a dishwasher would be. And so maybe humanoids, I love them. Unfortunately, excitement in the future isn't a zero sum game. I can love custom made robots. I can still be excited about humanoids as well.

• 29:45Harry Stebbings?

  And then why would you, what's the bull case to be excited about humanoids then? Cause I just listened to you there and I'm like, it's all completely rational, right and fair. Paint the bull case.

• 29:53Speaker 02

  I think the bull case is at home, like when it actually the speed doesn't matter, but there's just a whole host of many different tasks. So you can have a Roomba and the Roomba will do one thing better than a humanoid. You can have a dishwasher and will do that thing better than humanoid. But if you now also have like fifty other things, sort my socks and do this and open the door and get a package from outside, put it inside and like fifty or a hundred other little tasks, none of which you actually have to at scale massively every day, all the time. So the custom robot form factor doesn't make sense. Then I think humanoids can be helpful. And if they're quiet enough and the task execution is quiet enough, you can have them work slowly all night, right? You have a party and you wake up and the whole place is clean again.

• 30:36Speaker 02

  Now, the problem is that that is, and it's like these cluttered environments that are all different. There's no standardization. That's really, really hard for AI.

• 30:44Harry Stebbings?

  Also, the dexterity that's required within the hands to know about picking up a glass and how much tension to put on it versus not breaking it to plump a pillow, as stupid as it sounds.

• 30:54Speaker 02

  Yeah.

• 30:54Harry Stebbings?

  I was speaking to one founder the other day that was so far away from that.

• 30:57Speaker 02

  It's, it's similar to what we see and what we have seen with like radiology, for instance. It's very easy to build one radiology classifier for one thing and make that better than a human. But then there's this very long tail of things. And so you need a lot of money and a lot of resources and a lot of data to see the long tail of all things in radiology before you could actually automate a radiology process fully from end to end and get it fully FDA certified and so on. And in household robotics, you have the same very long tail.

• 31:28Harry Stebbings?

  It's so interesting. It's almost like the opposite of L L Ms where like the technology is ready and it's actually human adoption that's the thing that's stopping it versus like robotics, which is that actually the adoption, I think, would be there to have a cleaner in every household. But actually it's the technology that's blocking the adoption.

• 31:42Speaker 02

  It's still, it's still like a lot of hard work and very interesting research and more and more development. I think like Facebook Meta had open sourced like some new finger tips also that had some pressure sensors and stuff and you'd have to incorporate that. I think it's doable now. Uh, it's just like, it's a lot, a lot of work.

• 32:00Harry Stebbings?

  I've had guests on the show before. Talk about different medical discoveries. And my mother's got MS. You mentioned about biology. Do you think that we will find the solution to some of the biggest medical problems in the next ten years through some of the discussion that we've had already?

• 32:16Speaker 02

  Yes, that is that is one of like several like several chapters in my book are about that. I think that's one of the most exciting part of AI that AI will change. In many ways, science, I think, has been stuck in understanding the micro really, really well. But has not yet found a tool to understand complex systems really well. Like, we know every neuron and how every single neuron in your brain works. Just from first principles, you can take one out and it has all the synapses and you can compare and then you see, okay, there's inputs, outputs, and this is how it fires. But then when you put a bunch of them together and suddenly you have intelligence, like no one understands sort of these thresholds, these complex systems emerging and having these emergent properties.

• 33:00Speaker 02

  And, you know, why is that when the brain is slightly bigger and has enough neurons and the right setup, now all of a sudden you have true like human intelligence versus like much less strong like animal intelligence, but they can have better visual intelligence, but not language intelligence and like all these different things, like these emerging properties, same of like microbiome in our gut, we understand like how one bacterium, like what one bacterium does. But then you put them all together, no one can really tell you like why like certain foods will have a skin, create a skin issue or change your mood and and things like that and how it all connects your microbiome. People literally do poop transplants. Because we don't understand and sometimes they work and work magically well and cure

• 33:41Speaker 02

  actual diseases for people because now they have better gut microbiome. And so I think there's so much complexity and AI is the perfect tool to tackle that kind of complexity because we can now have similar things. We understand how one neuron works. But when you have enough of them, you scale it up enough, all of a sudden you can have a conversation and people think it's a it's a human being on the other side. Right? And so AI is the perfect tool. Medicine is the perfect application for that.

• 34:04Harry Stebbings?

  Where are you most concerned about AI?

• 34:06Speaker 02

  Job changes are brutal in the moment. There will add a lot of pressure on social systems. I think long term, I'm an optimist, but short term, you've got to have good social systems and help people with a path in this new AGI future as their jobs change and fall away. The modern day ludites are illustrators, right? Because it used to be that you can charge $500 for an illustration. And if you cared about just seeing illustrations as art and you want to see as much art as possible in the world, you're excited about having I now create Ghibli and all other kinds of beautiful illustrations. But if that was your life's income and you spend years honing that skill and now that skill is just not worth it anymore, then it's cost five cents or less to make an illustration, you hate the technology, right? And this is understandable.

• 34:50Speaker 02

  Now, of course, just like the past day ludites, we all appreciate having a T-shirt and everyone, even in Africa, can have as many T-shirts as they want and it's very cheap. And we like, most humanity likes the outcomes, but most people don't like it when they get paid by the hour for those outcomes. And so that is a negative downside that I think it's on to governments to help their people not to block the technology and enforce not using it, but actually use it and use the proceeds and help people to learn new kinds of skills.

• 35:18Harry Stebbings?

  Everyone always says, well, you know, we've always had this before. You've got the agricultural revolution, you've got like bunting PCs into entering workforces in the 90s. That took actually multi-decades. It took long, long time periods to actually move physical machinery into large farms. And in still, a lot of cases, there isn't in some parts of the world. Same with PCs, it took a decade actually for PCs to fully be, this is a software update. So the speed of transition is completely different.

• 35:41Speaker 02

  Yes and no. It is a software update, but you'd be surprised, and we talked about this a little bit earlier, like the adoption is not as fast as you think. It takes people to change their process, it takes them some time. I don't think it'll happen as quickly as people think. There's still, I think 60% of all adults in the US have never talked to a chat model, like at all. And that's the US. I can just go to Europe, I'm sure that is even higher.

• 36:04Harry Stebbings?

  Do you think in the future, we will choose which model we use? I find this archaic that we are choosing the model when we're putting in a prompt. It's like, it's like taking your car to the, you know, car mechanic and being like, I want that span and not that span. It's like, no, you'll just get given the best model for that specific prompt and request, no?

• 36:22Speaker 02

  Yeah, yeah. So it's actually something that we have shipped internally and I think it'll come out in a week or two on u.com, like to just automate that orchestration completely away. And only the power users who really want to can double click into the interface and then choose which model to use.

• 36:36Harry Stebbings?

  Find that absolutely bizarre.

• 36:37Speaker 02

  Hundred percent.

• 36:38Harry Stebbings?

  Staying on that, they're like, what is the right solution for government? Everyone says UBI. This will lead to mass unemployment in in only the short term. And retraining is hard.

• 36:48Speaker 02

  Retraining is hard. I used to be a big fan of UBI and I thought that seems like a fair thing to do. And I think you have to, one, acknowledge that everything falls in some normal distribution. I think that we have to teach kids a drive to want to create something, like to actually have desires to make improvements to the world. In fact, I think one of the biggest civilizational unlocks would be to all agree that entropy or darkness in the universe is the enemy and that we should like spread consciousness into the universe. If we can all agree that's the goal, then we can always be motivated to do more of that, to to spread intelligence and intelligent entities into the universe. That would be beautiful.

• 37:31Speaker 02

  But zooming back in, I think the problem is that while most people complain about their jobs, it does give them meaning. It does give them meaning to be a valuable part of society and to have earned something that they can then give to their family, to their kids and so on. And so I think UBI will take that meaning away. And we're already in a meaning crisis with the technology and and society that we have set up for ourselves in many places, and I think that will just exacerbate it.

• 37:56Harry Stebbings?

  Do you think, speaking of like meaning, do you think we will have AI friends in a more significant way than we have human friends?

• 38:03Speaker 02

  I think we will have more significant AI friends, but I hope it's not more significant than our friendships with people.

• 38:10Harry Stebbings?

  And you believe that we'll have like the AI companion?

• 38:12Speaker 02

  You know, I've been an investor in like Replica like many, many years ago. To me, you can find beautiful examples where people just say, look, I used to journal. Now I write into this thing and sometimes it asks me questions back and it makes me feel like some someone cares. Think about your best friend. If they kept coming into you every single day with all their little problems, you know, like that problem didn't cross my threshold of being relevant enough for you to tell me about it, you know, like, but some people have like that desire to just tell other people about their problems all the time. And very few people have the just time and capacity and emotional capacity to hear that all the time, every day. But so I think it's a perfectly fine use case for people to work through their problems, just like they used to write a diary.

• 38:57Harry Stebbings?

  So if you were advising a nineteen year old or a twenty one year old coming out of university or school about where they should focus to prevent themselves bluntly being in a position where they have to retrain, where would that be?

• 39:10Speaker 02

  It's similar but different, like, cause in high school, after you finish high school, you can still study for for a couple of years. So that's a that's a big difference. Once you're already out of college, it's like it's much harder. But knowing computer science is incredibly important, like knowing how to program is incredibly important.

• 39:26Harry Stebbings?

  Why do you say that with the commoditization, a lot of low level programming?

• 39:30Speaker 02

  For the same reason why we're speaking English. And like, I could have stayed in Germany and never learned English. And then every time you say something, I, you know, for a while, I would type it in and then hear it. And then like every, you know, two minutes, we'd have like some useful bit of information going back and forth. Or eventually I'm going to have just like a full bubble fish in my ear and it will translate it hopefully fairly well and get the connotations and maybe capture my voice. And maybe that technology is like fully there in five years or so, like so, so, so good, so cheap, so prevalent that everyone will just travel with it and everything. I think even then it'll be useful for some people to know some languages. But do we need to translate as much?

• 40:13Speaker 02

  No, actually, in my masters, I was studying Chinese back in Germany. I just fell in love at that time with statistical machine learning, pattern recognition and statistics and what we now call AI. And like, realized like that is a more useful way to spend my time than learning five more languages, which is beautiful for me, but not useful for humanity. And so I do think that will fall away. But programming isn't just about programming itself and being like.

• 40:39Speaker 02

  IT or being a programmer or a developer or so on. It's also about a different way of thinking and it's a different way of understanding the world that you're in. And so if you have an understanding of that, then it feels less like magic. And you are more empowered to actually contribute to that world. So I think that's important, even if you want to study law, medicine, chemistry or whatever. You should combine it with computer science because all of these fields are going to change over the next couple of decades.

• 41:06Harry Stebbings?

  How do engineering teams and their structures change over the next couple of decades?

• 41:11Speaker 02

  I think the biggest problem and biggest worry I have is that the entry level jobs right now are more and more automatable. And so you hopefully have companies that have a long enough time horizon such that they're willing to train people. Even though any I could do the job because they're either cheap enough or like they just have a long term view on the world. And then by the time they are senior enough, hopefully I hasn't automated that senior job also. And then they can be in that senior role and then manage all of the AI agents because they understand the processes. Turns out if you've never done something, it's much harder to manage someone to do it for you. Same with like vibe coding, right? It's like, it's super fun. It's awesome. I just posted this really funny video. This comedian about vibe coding, but if you don't know how to program at all.

• 41:56Speaker 02

  You're also not going to be as good of a vibe coder. Like there's just certain things like complexity theory. Like you just say like, oh, do this sorting for me or do this. You know, like if you don't know how sorting algorithms work and like how fast they could be like make it faster, make it faster. And then sometimes like there are certain complexity theoretic upper bounds of how fast a sorting algorithm can be. And you just telling it that it'll go to jail if it doesn't do it faster, like is not going to make it any more, any faster. And so it's useful to understand that you also have seen that even L M's benefit in their general reasoning skills. When they know and have seen programming as training data, and I think that's true for people too.

• 42:32Harry Stebbings?

  So will engineering teams be smaller?

• 42:34Speaker 02

  I think so. I think a lot of teams will be more efficient. We'll be managers of AI. I think all the boring bits of all work. I think one of the ways to think about the future of work is, one, we'll all become managers of AI. Number two, managing is hard, and so we have to train people to it. What that means is like, in the future, every time you do a job, like a dozen, two dozen times, you're gonna be like, wait, why hasn't the AI now taken that over from me? Why do I still have to do this repeatable boring thing? I personally hate repeatable boring things, so I'm all for it. But that is a mindset shift that schools don't teach yet. And that will take a generation of kids to have to grow up in this mindset.

• 43:10Harry Stebbings?

  We see this kind of continuous battle between windsurf and cursor. And a lot of people say that there's actually very little switching cost between moving between the two. How do you think about that and how that market plays out?

• 43:23Speaker 02

  Yeah, and Codium, I wish I could have invested in cursor. The, one of the founders was actually an intern at U.com, and I would have loved to invest. But yeah, so, and then Codium, I fortunately was able to invest in.

• 43:34Harry Stebbings?

  Was he brilliant?

• 43:35Speaker 02

  Yeah, he was very, very smart. I really tried to like keep him at U.com and stuff. He went from intern to, I think, CEO, yeah, like, and of cursor, and so, yeah, brilliant.

• 43:43Harry Stebbings?

  Why did you not, why did you not follow his money while I, but I try and invest when he started?

• 43:47Speaker 02

  I tried. Some other investors were like, oh, well, they had an even closer relationship with him.

• 43:52Harry Stebbings?

  Bastards. Hate them bastards. Okay.

• 43:56Speaker 02

  I'm happy for him. So, yeah, there is very little switching cost. Same is true for LMs in the consumer world, right? They're not, none of them are that amazing yet. None of them do enough personalization yet. Now, where there is a lot of switching cost is if you have company internal data or you have unique data assets.

• 44:11Harry Stebbings?

  How concerned are they for you in terms of giving very, very private data to you? Where traditionally it might be held on-prem, they're nervous about security, they're nervous about privacy. How much of an actual barrier is that?

• 44:23Speaker 02

  It's a huge barrier. LMs are garbage in, garbage out to a large degree. So if you have a search model or an index and you ask like, what's new with Trump? And that search index brings back pages from like six years ago. The LM will tell you wrong and outdated things about that query. So. The search is kind of the often forgotten infrastructure layer for L M's. And so companies are rightfully concerned about privacy and all of that. And that's why we have zero data retention. We have agreements that we don't train any model on company internal data. That's how you get into enterprise.

• 45:01Harry Stebbings?

  Do you think we are seeing the biggest corporate misbehavior from this generation? Taking chat G P T and putting company data in it that they shouldn't?

• 45:10Speaker 02

  They should stop doing that if they have done it in the past. Yeah.

• 45:13Harry Stebbings?

  Do you not think they will?

• 45:14Speaker 02

  I I think a lot of people are doing it, but we also know now we have customers who are like, okay. Like, we want to know exactly which models are accessible and have these security requirements and so on. And that's where you really can't be like a great consumer company and a great enterprise like company at the same time. It's very, very hard and we're focused on on that enterprise and make sure that the security is there, the trust is there, the accuracy of the answers there, the ability to say, I don't know, is there? All of these things are are important aspects.

• 45:43Harry Stebbings?

  To what extent is cash and the weight of cash a weapon in this market?

• 45:48Speaker 02

  In terms of like startups?

• 45:49Harry Stebbings?

  Like, is cash a moat for L L M's and for companies in this market?

• 45:54Speaker 02

  It can be, but we have seen some companies now that had raised hundreds of millions of dollars and still died. We're going to see a correction when companies are trading hundred eighty X. They're A R R and they don't have a real moat. And like there's the switching cost is close to zero to go to deep sea or something else like.

• 46:15Harry Stebbings?

  Have you seen the retention numbers on deep sea?

• 46:17Speaker 02

  I have not.

• 46:18Harry Stebbings?

  They're pretty shit.

• 46:19Speaker 02

  I'm not surprised.

• 46:20Harry Stebbings?

  No one's staying with deep sea.

• 46:21Speaker 02

  I'm not surprised. Yeah.

• 46:22Harry Stebbings?

  And so it's like, actually, is that very valuable? Does that not prove that the ultimate value in this market is consumer brought?

• 46:28Speaker 02

  In some ways, I think A I has has gotten so exciting for so many people that the startup world is going back to the basics. Like, you know, it's just like when there were a hundred photo sharing apps and only one Instagram and maybe a Flickr and so on. Uh, no pun intended, Flickr, like. I think we'll see something similar in AI. It just goes back to is your branding good, your marketing, your sales, your distribution? And then, of course, a lot of the technology things, but those get commoditized more and more, just like sharing photos was a fairly commodity capability. But a lot of little subtle details were better for Instagram. Now, that's consumer. In consumer, we usually end up in monopoly or duopoly situations. Enterprise is a very different world.

• 47:09Harry Stebbings?

  To what extent do you think we see ChatGPT and OpenAI move and kill a ton of different, more consumer-facing apps? So we looked at a company recently and they were in the, basically, picture creation space for fashion. So you take a picture of a model in a T-shirt and it'll do it in all the T-shirts in the SKU line.

• 47:26Speaker 02

  Yeah. Yeah.

• 47:27Harry Stebbings?

  OpenAI just released that the other day. There's ten of them that just died. Background remover, ten just died.

• 47:33Speaker 02

  Yeah.

• 47:34Harry Stebbings?

  To what extent will we see OpenAI kill a generation of companies of this material?

• 47:39Speaker 02

  I actually don't think they're all gonna just die overnight. Like, think about speech recognition. It's very commodity. There's tons of open source speech recognition, like, algorithms you can download and so on. There's still several companies that make millions and like tens or even hundreds of millions of revenue doing speech recognition just perfectly. That last little bit, like, you don't just throw like a full feature-length movie and make it in this new style, like, with ChatGPT. Like, companies are gonna wanna have custom solutions for them. And even like that fashion thing, I think there's a space where you take the picture and like it's just integrated in your camera, in your workflow.

• 48:21Speaker 02

  You can change the SKU, you can change the model's hair color and every little pixel is just perfect. So you can put it on a big billboard. I think there will still be like specialized companies that go and do deeper things than you could do if you knew how to use it all yourself and like you're really clever and like you are an early adopter.

• 48:39Harry Stebbings?

  What do you think is the biggest misconception that people have today around AI and LLMs?

• 48:44Speaker 02

  I think the biggest misconception is that we we have sort of this bimodal like black and white when things are often in some gray in between. Like, some people think it's gonna take off over all the jobs. And then because it's so brilliant, like overnight, like everything will be gone. And then like the next night it will kill us all, right? Like, it's just like this extreme optimism that then can sometimes also switch into extreme like pessimistic optimism of like, oh, the technology is so good and becomes fully self-aware and conscious that it will then obviously want to kill us all. So there's like the misconception on that side. Then there's the misconception to still like, and I see this in Germany still, there's still people like, ah, you know, two years ago was crypto, this year's AI, like, maybe it'll just go away. And like, that's also a huge misconception that still exists in the world.

• 49:26Speaker 02

  It's hard for you and I to imagine as we live sort of in a bubble, but I see it when I travel.

• 49:30Harry Stebbings?

  Do you think with this advancement, we have escaped Moore's law in terms of our speed of progression? Is Moore's law ever escapable?

• 49:37Speaker 02

  Of course. And like, you know, I think parallelism has helped a ton, like Nvidia shows us that, yeah, we might not double the number of transistors or whatever every eighteen months, but like we may have more parallel ones. And then there's still the the black horse of quantum, which may, you know, a lot of announcements, unclear if they're like how real they all are, but like that will also be a major shift when it finally does happen, but it might still take five, ten, fifteen years.

• 50:01Harry Stebbings?

  What would be the most significant changes that result from quantum development in a way that we would like to see?

• 50:07Speaker 02

  I mean, there's the obvious one that we don't want to see, which is like, all the passwords need to be re-encrypted and changed. And all the data that has been leaked but is encrypted might get decrypted. And hence, like, people know what's in some things of the past that they had hacked, but couldn't really decipher yet or decrypt. I think on the positive side, what I'm really excited about is getting quantum computers to a scale where we can simulate cell. We can right now, like, really, really accurately only simulate like a few hundred atoms at best and how they really truly interact with one another. And like neural nets can approximate that. There's a lot of cool things where you can hack simulations, but to from real first principles, like really deeply.

• 50:51Speaker 02

  Model physics and chemistry that is complex and eventually biology like that will be such a massive unlock because in AI, anything you can simulate, AI can solve every problem that domain. Like you can simulate Go or chess, like you can simulate a computer game because it's in the computer and you can play around with it. Obviously, AI is going to solve that at some point like that. None of those like results were that surprising to me. Though they again amazing like marketing and a grazing like actually doing it, right? It's it's really hard, but it's not that surprising. But what we we can't simulate most of the interesting things in the world such as a cell. But once you can simulate a cell and then multiple cells and organs and organisms like all of a sudden, AI can try billions of different things on how to cure that cancer, how to cure MS, how to cure like

• 51:36Speaker 02

  all the bacteria and viruses and like all these different things like it will be such a massive unlock to be able to simulate those things with quantum computers and maybe also more and more with normal computers.

• 51:46Harry Stebbings?

  Dude, I would love to do a quick fire. I've peppered you with different questions, literally from every different spectrum. And I'm still I'm still thinking about the negative consequences and what kids need to do. But yeah, well, let's do a quick. What what did you believe that you now no longer believe?

• 52:00Speaker 02

  Like I said, I think a big one for me was this sort of skepticism about the future. I'll tell you a story like one of the co-founders of OpenAI and I started a bet, I think, seven years ago or so, where he said, we'll have AGI in like nine or 10 years. And I was like, I mean, I'll work hard on research to make my prediction be wrong, but I I really don't think we will. This was at like an AI conference and we're both like still more more junior than we are now. And he had this like really strong belief and you know, they worked on robotic hands and they're like another step towards AGI. And I'm like, I mean, it was a cool robotics project. They worked on Dota game playing at OpenAI and I'm like, that's a cool RL project. And they're like another step towards AGI.

• 52:41Speaker 02

  And then, you know, they saw like our prompt engineering paper, which they cited and like like that that NLP route was a more legit step towards AGI than than the previous things. But long story short, we we did this bet and in the bet, he has to win. I think it ends in 2027. So three things have to be true. We have to have a personal robot. That cleans the whole house the way my cleaning team does and it needs to be purchasable like for reasonable amounts of money. It needs to solve a millennia math problem and it needs to like translate a book perfectly so that the actual author would be like, that should be my official translation. All three have to be true for him to win the bet and I will probably still win my thousand dollar bet, but he became a billionaire in the process of proving me wrong.

• 53:25Speaker 02

  And so that kind of shows you, you have to just have a constructive optimism and that was probably something that a belief I changed. It's just like, even if you don't think it can be quite possible, you should try to make the most audacious goals and set the most audacious goals for yourself. And then basically work on pragmatic milestones towards those goals.

• 53:46Harry Stebbings?

  OpenAI at three hundred, Anthropic at sixty or Grok at fifty, which would you most invest in?

• 53:52Speaker 02

  Can I choose none? I mean, open source puts a lot of pressure on it. It's just.

• 53:56Harry Stebbings?

  You really, how do you think about the distribution of value between closed source and open source and do you not think it follows the tradition by the the latest of Linux?

• 54:03Speaker 02

  Like if you're really sophisticated, you can use the open source more and more. And as the open source models have caught up, it's just very hard to say this is like super unique technology. It just becomes like, you know, for a while in technology, like having a really good database was really powerful. And indeed, there was like, there is an Oracle now. And so if you're really, really crazy large scale, then you might still use an Oracle. But for a lot of other folks, they can just use simpler, other, smaller databases and you have that wild horse of.

• 54:33Harry Stebbings?

  Do you think they will though? It's like you said, they're the final ten percent for, you know, speech recognition, which is like, yeah, they can do, but they don't.

• 54:40Speaker 02

  Yeah, I think like, again, it's very different for consumer, like OpenAI actually has a massive number of consumers and that's super duper powerful. And so, yeah, all of them have advantages, like Sonar three point seven is actually the best model for a lot of things, especially in coding and engineering and so on. Like you said, it's a little bit unclear to me. I I personally would just, I I also just love investing in early stage where you can have thousand Xs and so on. I just don't see thousand Xs for for those companies.

• 55:06Harry Stebbings?

  You care about money.

• 55:07Speaker 02

  I don't.

• 55:08Harry Stebbings?

  Has that always been the case?

• 55:09Speaker 02

  It has been always the case. Yeah, I was like, I was an academic for a long time. I want to be professor. I want to do research. I miss the research also now still. I now have to, I realize at some point you have to care about money because money is one of the best indicators of impact and allows you to do epic, cool things, but I don't intrinsically care about it.

• 55:31Harry Stebbings?

  How do you think about defining or measuring success for yourself?

• 55:35Speaker 02

  I think a lot of it boils down to sort of how much positive impact you've had in the world that matters enough so that people in the future will still remember it.

• 55:44Harry Stebbings?

  Is there an element of ego to that, about being remembered as the creator of?

• 55:48Speaker 02

  It's undeniable. Like, in some ways, it's, it's beautiful if you do it, but the difference here for me at least is that I'm okay just being remembered by the people that know. Like, most people have no idea who invented penicillin. And most people can't name the people who invented the cure for their cancer. Because they're just like, some doctor gave it to me and I got this medication now, right? But the people in the know know, and that would be good enough.

• 56:15Harry Stebbings?

  What trait are you slightly ashamed of but has contributed to your success?

• 56:20Speaker 02

  I do get quite extreme too, if I get into, into a zone and then like everything is kind of a nuisance. And, and then you just get like just really intense about something. No, I'm also happily married, so that helps not being lonely. But yeah, I don't, I don't need people all the time around me. In fact, I'm an introvert. I enjoy, I have like sometimes events at my, my, my place of like hundreds of people and like I enjoy that a lot. I get sort of, there's an activation energy and then you're like, all right, I'm now in this mode. But afterwards, I'm like, I don't have to see anyone for like another month now.

• 56:51Harry Stebbings?

  A final one, when you look forward to the next 10 years, what are you most excited for that you think is realistic and we will see happen?

• 56:57Speaker 02

  Improving longevity is, is like one of those AI plus bio corollaries or follows. That is just really, really exciting. And I think longevity is massively underrated. And there are a lot of people who are snarky and say, oh yeah, whatever. I don't care. I want to like, they, all people say that until it's a few days before they're dead and they're in pain and they're like, fuck, I wish I like was a little bit healthier, you know, beforehand. And was that really worth it to like drink all the time, not sleep enough and so on. So I am trying to be better about it personally.

• 57:31Harry Stebbings?

  What do you do that you'd most like to stop doing?

• 57:33Speaker 02

  Most like to like not sleep enough.

• 57:35Harry Stebbings?

  What's your sleep schedule?

• 57:37Speaker 02

  I'm traveling right now, so it's all screwed up. And that's like, that's actually one of the worst parts of of traveling now is like, it just messes with your sleep. Actually, I stopped drinking last year, also for longevity reasons. And life is like twenty percent less fun, especially in like some evenings and so on. But, you know, it is better. But I realized actually after like not drinking anymore, that the majority of times when I felt bad in the morning traveling wasn't actually from alcohol, it's just from like sleep deprivation and like shitty sleep schedules. So I I feel just as bad like.

• 58:06Harry Stebbings?

  I I totally get you. Final one, final one, I promise. The magic you want is in the actions you are avoiding. What actions are you avoiding?

• 58:15Speaker 02

  I I I like whenever I see something that I should be doing, I try to write it down and then I try to work towards it. So I I like to think I don't avoid many actions. But I guess when it comes to longevity, like every night I wake up in the morning and I'm like, I should have gone to bed earlier last night and I would probably feel better right now. So maybe that's.

• 58:34Harry Stebbings?

  Is the EU fucking itself AI regulation?

• 58:37Speaker 02

  Unfortunately, the EU has shot itself in the foot. A lot of different kinds of regulation and in particular AI regulation destroying a fledgling ecosystem that could never get off the ground.

• 58:47Harry Stebbings?

  What should we be doing that we're not doing? What would you do if you were in charge?

• 58:51Speaker 02

  A lot of things. I would make computer science a mandatory subject in all schools. A very sought after minor for almost every major in college. I would help people and excite people more. It's also a marketing thing to some degree and just mindset shift for people to start their own companies. Because every person that cares about outcomes and outputs loves AI. It's only the people who think of how to make money in terms of our spent that might not like AI. And so giving, trying to do some marketing for the populace to have that mindset shift of like, I can make outputs. I can be an owner and then creator. And then possibly start a sovereign wealth fund so that, you know, you can just participate in all the upside

• 59:36Speaker 02

  of amazing AI technology worldwide, not just like sovereign wealth fund that only invests internally inside the country, but also. Externally, I'd make it easier for large enterprises to buy small startups also, maybe get some tax benefits or something for it. Because there's a lot of not invented here mentality in large European companies. And so you just don't have as much of that ecosystem. I would reduce the barriers to go public so that you have even more benefits for startups and have, you know, they're basically two exit routes, right? You either get acquired or you go public for VCs in the ecosystem. And so you'll get more VC if both routes work out better. I'll stop there, but been asked that question a lot.

• 1:00:19Harry Stebbings?

  My god, that was fantastic.

• 1:00:20Speaker 02

  I try to be helpful when people ask me that question.

• 1:00:22Harry Stebbings?

  Dude, thank you so much for putting up with me, completely just peppering around. Like, as I said, I have these quite home schedules and then I'm like, fuck it, let's just go for it. You've been fantastic. So thank you so much.

• 1:00:33Speaker 02

  It's always a pleasure talking to you, man.

• 1:00:36Harry Stebbings?

  That was so much fun to have Rich in the studio there. And if you want to see the full episode, you can find it on Spotify by searching for 20 VC. But before we leave you today, Secureframe empowers businesses to build trust with customers by simplifying information security and compliance through AI and automation. Thousands of fast growing businesses, including NASDAQ, Angel List, Doodle and Coda, trust Secureframe to expedite their compliance journey for global security and privacy standards, such as SOC 2 and ISO 27001, CMMC, MIST standards and more. Backed by top tier investors and corporations like Google and Kleiner Perkins, the company is among Forbes' list of the top 100 startup employers for 2024. G2's best software awards for highest satisfaction products

• 1:01:20Harry Stebbings?

  and a recipient of the 2024 Cybersecurity Excellence Awards, something I definitely never got in school myself. Learn more today at secureframe.com. And speaking of trust and intelligence, let me tell you about Harmonic. Did you know that half of the 27 companies started last year by OpenAI alumni are still in stealth? I discovered this on Harmonic, the complete startup database used by Excel, Insight, Menlo and hundreds of other leading VCs, as well as go to market teams from the likes of Notion, Brex and Google, to find the best startups and founders, even In stealth, few things annoy me more than missing a round where we know the founder and should have led the round. But now, this is a problem of the past.

• 1:02:01Harry Stebbings?

  Harmonic maps everyone your team has ever met, emailed or connected with to the source of truth for startups. So when that company that was just a little too early suddenly gains traction, you won't be too late. Or when you find the perfect company for your thesis or product, you'll know that Sarah just happens to know their CTO. How does Sarah know everyone? Who knows? She's a genius. Learn how VCs and GTM teams find the best startups six months ahead of the competition on harmonic.ai. And while Harmonic helps you stay ahead of the curve in startup investing, what if your phone could help you do the same with your wallet? We spend nearly half of our waking lives glued to our phones, upwards of 50 hours every week.

• 1:02:41Harry Stebbings?

  Recently, one company transforming this reality stood out so much I personally became a shareholder. Mode Mobile. Mode Mobile created the Earn Phone, a smartphone that pays you for daily activities. Instead of big tech profiting billions from our attention, Mode returns over 325 million directly to users through earnings and savings. Mode's revenue surged an incredible 32,000% in three years, recognized by Deloitte as 2023's fastest-growing software company in North America. And here's why I'm excited. Mode's equity offerings have raised over 30 million from 20,000-plus retail investors, one of 2025's standout public raises. You can now join me as a shareholder with as little as $1,000 at invest-modemobile.com/20vc.

• 1:03:26Harry Stebbings?

  For a limited time, unlock up to 100% bonus shares and a free Earn Phone. Email us for investor brief at [email protected] or check out invest-modemobile.com/20vc. As always, I so appreciate all your support and stay tuned for an incredible episode coming on Monday with the founder of Dave, the most incredible turnaround in the public markets. They went from over $1 billion to a $50 million market cap. The turnaround story on Monday.